incubator-cassandra-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Rishi Bhardwaj <khichri...@yahoo.com>
Subject Re: Atomic Compare and Swap
Date Mon, 21 Jun 2010 21:19:11 GMT
I have read the post on cages and it is definitely very interesting. But cages seems to be
too coarse grained compared to an Atomic Compare and Swap on Cassandra column value. Cages
would makes sense when one wants to do multiple atomic row, column updates. Also, I am not
so sure about the scalability when it comes to using zookeeper for keeping locks on Cassandra
columns... there would also be performance hit with an added RPC for every write. I feel Cages
maybe fine for systems when one has few locks but I feel an atomic CAS in Cassandra would
help us avoid distributed locking systems and zookeeper in many other simpler scenarios. For
more complicated (transaction like) things, using Cages may be fine. Then again doing a read
before write for CAS in cassandra will make CAS at least as slow as a read, which I believe
will still be better than taking a single column lock from zookeeper.

What do other folks think in this regard? From whatever I have read, I believe CAS is feasible
in Cassandra without hurting the normal write path performance. Only for CAS writes would
we have to pay for the read before write penalty. I am going to do feasibility study for this
and would love any pointers from others about this.

Thanks,
Rishi



________________________________
From: Rauan Maemirov <rauan@maemirov.com>
To: dev@cassandra.apache.org
Sent: Mon, June 21, 2010 11:27:02 AM
Subject: Re: Atomic Compare and Swap

Have you read this post?
http://ria101.wordpress.com/2010/05/12/locking-and-transactions-over-cassandra-using-cages/
I guess, you will like it.

2010/6/22 Rishi Bhardwaj <khichrishi@yahoo.com>

> I am definitely interested in taking this work up. I believe the CAS
> functionality would help in a lot of different scenarios and could help
> avoid use of other external services (like zookeeper) to provide similar
> functionality. I am new at Cassandra development and would really appreciate
> pointers from the dev. community about how to approach/start on this
> project. Also how feasible is the approach mentioned below to implement the
> CAS functionality? It would be great if we could have a discussion on the
> pros and cons.
>
> Thanks,
> Rishi
>
>
>
> ________________________________
> From: Sriram Srinivasan <sriram@malhar.net>
> To: dev@cassandra.apache.org
> Sent: Sun, June 20, 2010 9:47:37 PM
> Subject: Re: Atomic Compare and Swap
>
>
> I too am interested in a CAS facility.
>
> I like Rishi's proposal. One could simply use a version number as the
> logical timestamp. If we promote CAS to a consistency level, it would rate
> higher than a quorum. One pays the price for a more complex write path to
> obtain the requisite guarantee.
>
>
> On Jun 21, 2010, at 4:03 AM, Rishi Bhardwaj wrote:
>
> >
> > Heres another thought I had, if say the user always wrote with quorum (or
> to all) nodes then can't we implement CAS (compare and swap) assuming that
> user employs logical timestamp and Cassandra doesn't allow writes to a
> column with same or older timestamp. Here's the scenario I am thinking
> about:
> > Say we use logical timestamp for a column value and lets assume the
> current timestamp is t. Now say two clients read this column and generate
> concurrent CAS (compare and swap) operations on timestamp t and for both the
> writes the resulting new timestamp would become (t+1). Now if we don't allow
> writes to a column with same timestamp then only one of these writes would
> succeed. Of course another assumption is that if a third CAS write with
> compare on logical timestamp (t - 1) came in, that would be denied as I
> believe Cassandra doesn't allow "older" writes to win over "newer" writes.
> Do you think such a thing can be accomplished?
>
>
>
>



      
Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message